Hi there,
Thanks for the great work! Since CodeActInstruct is all multi-turn, I wonder what's the specific formatting of the data when passing in for SFT? Specifically, for one entry of the dataset, do you put all but last response (assistant response) as input and just the last response as output? Or do you do split by turns (for the example instead of passing in 1 entry I would have 2 SFT entries where the first entry is conversation up until first assistant response, second entry up until second assistant response etc.)
Hi there,
Thanks for the great work! Since CodeActInstruct is all multi-turn, I wonder what's the specific formatting of the data when passing in for SFT? Specifically, for one entry of the dataset, do you put all but last response (assistant response) as input and just the last response as output? Or do you do split by turns (for the example instead of passing in 1 entry I would have 2 SFT entries where the first entry is conversation up until first assistant response, second entry up until second assistant response etc.)